The role of linguistic and emotional information in learning and representing the meaning of abstract words

• Applicants: Dr. Laura Bartylla; Professorin Dr. Wiebke Petersen
• Subject Area: Biological Psychology and Cognitive Neuroscience ; Applied Linguistics, Computational Linguistics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 563106273

This project integrates cognitive psychology and machine learning to investigate how humans and language models represent the meanings of abstract words. Within the framework of grounded cognition theory, it explores the roles of linguistic and emotional experiences in shaping word meanings. While concrete words rely on sensorimotor grounding, abstract words, like “love” or “conflict,” depend more on internal (e.g., emotional, interoceptive) and linguistic experiences. Language models encode statistical patterns in text, capturing linguistic and emotional nuances. Despite differing mechanisms, human and machine representations show parallels, providing a basis for systematic comparison. The project aims to uncover how linguistic and emotional information shapes meaning representations, therein comparing humans and machines. It will explore how linguistic and emotional properties influence semantic features generated by humans, the neurophysiological stages of processing where experiential grounding occurs, and whether emotion-informed language models are more human-like in their representations. To address these questions, the project integrates psychological experiments with machine learning across four working packages. First, human participants will produce semantic features for real abstract words and their brain activity is recorded using electroencephalography during word processing. Behavioral analyses will classify experiential features, while neurophysiological analyses will identify when linguistic and emotional grounding occurs. Second, language models will be trained on human-generated semantic features, comparing standard models with emotion-informed ones. Multimodal models combining EEG and linguistic data will examine how neural and vector-based representations align. Next, humans will learn novel abstract words via association with controlled linguistic and emotional inputs, enabling direct testing of how these dimensions shape word meanings. Finally, language models will be trained on the same inputs, allowing unprecedented comparisons between human and machine representations to reveal shared and distinct mechanisms. Led by an interdisciplinary PI team from neuroscience and computational linguistics, the project will provide empirical evidence on the experiential grounding of abstract concepts, identifying how linguistic and emotional dimensions contribute to meaning representations in humans and machines. Insights from cognitive psychology will enhance machine learning models, while computational approaches will inform psychological theories, fostering an interdisciplinary synergy to advance understanding of meaning representation.

• DFG Programme: Research Grants